Robust Bathymetry prediction from Satellite Gravimetry via Noise-Detecting Convolutional Neural Network

The inversion of seafloor topography from satellite gravimetry has been significantly advanced by neural networks (NNs), which leverage their superior nonlinear modeling to enhance reconstruction accuracy. However, these models typically depend on shipborne single-beam/multi-beam soundings as labeled training data, and their data-driven nature renders them vulnerable to label noise—a pervasive issue in marine surveys. To mitigate this challenge, we introduce a Noise-Detecting Convolutional Neural Network (ND-CNN). By dynamically recalibrating sample weights during training, the ND-CNN intrinsically identifies and prioritizes reliable samples while suppressing the influence of noisy ones. Experimental results demonstrate that our method substantially outperforms conventional CNNs, delivering robust performance even with 10\% label noise and a 21\% gain in prediction accuracy. The ND-CNN thus offers a powerful and resilient solution for seafloor topography mapping in scenarios where high-quality training data are scarce.